首页> 外文会议>International conference on artificial neural networks >IP-GAN: Learning Identity and Pose Disentanglement in Generative Adversarial Networks
【24h】

IP-GAN: Learning Identity and Pose Disentanglement in Generative Adversarial Networks

机译:IP-GAN:生成对抗网络中的学习身份和姿势分解

获取原文

摘要

Synthesizing realistic multi-view face images from a single-view input is an effective and cheap way for data augmentation. In addition it is promising for more efficiently training deep pose-invariant models for large-scale unconstrained face recognition. It is a challenging generative learning problem due to the large pose discrepancy between the synthetic and real face images, and the need to preserve identity after generation. We propose IP-GAN, a framework based on Generative Adversarial Networks to disentangle the identity and pose of faces, such that we can generate face images of a specific person with a variety of poses, or images of different identities with a particular pose. To rotate a face, our framework requires one input image of that person to produce an identity vector, and any other input face image to extract a pose embedding vector. Then we recombine the identity vector and the pose vector to synthesize a new face of the person with the extracted pose. Two learning pathways are introduced, the generation and the transformation, where the generation path focuses on learning complete representation in the latent embedding space. While the transformation path focuses on synthesis of new face images with target poses. They collaborate and compete in a parameter-sharing manner, and in an unsupervised settings. The experimental results demonstrate the effectiveness of the proposed framework.
机译:从单视图输入合成逼真的多视图面部图像是一种有效且廉价的数据增强方法。另外,它有望更有效地训练用于大规模无约束人脸识别的深度姿势不变模型。由于合成和真实面部图像之间的姿势差异较大,并且需要在生成后保留身份,因此这是一个具有挑战性的生成学习问题。我们提出了IP-GAN,这是一个基于对抗性网络的框架,用于区分面部的身份和姿势,以便我们可以生成具有各种姿势的特定人的面部图像,或生成具有特定姿势的不同身份的图像。要旋转脸部,我们的框架需要该人的一个输入图像来生成一个身份矢量,而任何其他输入脸部图像都需要提取一个姿势嵌入矢量。然后,我们重新组合身份向量和姿势向量,以使用提取的姿势合成人的新面孔。引入了两个学习路径,即生成和转换,其中生成路径着重于学习潜在嵌入空间中的完整表示。而转换路径的重点是具有目标姿势的新人脸图像的合成。他们以参数共享的方式在无人监督的环境中进行协作和竞争。实验结果证明了所提出框架的有效性。

著录项

相似文献

  • 外文文献
  • 中文文献
  • 专利
获取原文

客服邮箱:kefu@zhangqiaokeyan.com

京公网安备:11010802029741号 ICP备案号:京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有
  • 客服微信

  • 服务号